Palo Alto CEO Nikesh Arora demands 90% drop in AI token costs for mass enterprise adoption

AI Token Costs Create Major Barrier to Enterprise Adoption

Palo Alto Networks CEO Nikesh Arora has issued a stark warning about the future of enterprise AI adoption, telling CNBC that AI token costs must fall by as much as 90% before businesses can deploy the technology at scale 1

1

. Speaking on "Squawk on the Street" Thursday, Arora laid out a specific timeline: token efficiency needs to drop to roughly 20% of current levels over the next twelve months, and to just 10% by the following year 3

3

. His comments come as rising token costs have emerged as a major pain point for businesses, putting significant strain on AI budgets and making AI tools increasingly difficult for companies to implement.

OpenAI Efficiency Gains Fall Short of Enterprise Needs

When asked about OpenAI CEO Sam Altman's announcement that the company's latest model is 54% more token-efficient for agentic coding, Arora acknowledged the progress but made clear it represents only a starting point. "I think 54% is a good start," Arora said. "I think we probably need another turn at it" 1

1

. This token efficiency improvement, while significant, falls far short of what Nikesh Arora believes is necessary for large-scale enterprise AI deployment to become economically viable. Despite the current pricing challenges, Arora remains optimistic about demand. "The demand continues to be infinite, and as long as you have an infinite demand curve that you're facing, I think all these things will rationalize over time," he told CNBC 2

2

.

The Paradox Behind Prohibitive AI Pricing

Arora's plea highlights a genuine paradox in the enterprise AI landscape. While per-token prices have collapsed by 98%, total enterprise AI spending has actually tripled over the same period 2

2

. The culprit behind this counterintuitive trend is agentic AI usage, which calls models repeatedly to complete tasks. A single ambitious project can burn through massive resources, as evidenced by one developer whose agents ran up a $1.3 million token bill in just one month 2

2

. This means that cheaper headline prices don't automatically translate into lower costs—usage grows faster than prices fall, and bills continue to climb.

Token Shock Hits Major Enterprises

The strain from AI token costs is already changing corporate behavior across major companies. Uber exhausted its full-year 2026 AI budget by April, forcing Chief Technology Officer Praveen Neppalli Naga to say the company was "back to the drawing board" 4

4

. Chief Operating Officer Andrew Macdonald indicated Uber would weigh token costs directly against the cost of hiring engineers. This phenomenon, dubbed "token shock," has hit some of Silicon Valley's biggest spenders particularly hard. Companies including Microsoft have capped or restricted employee access to expensive AI coding tools after budgets blew past projections 3

3

. Some firms have moved toward cheaper open-weight models, including Chinese alternatives that are closing the gap with American labs.

Growing Executive Pressure on AI Vendors

Arora joins a widening circle of corporate leaders voicing concerns about what they see as prohibitive model pricing—costs high enough to keep AI from moving beyond pilots into genuine enterprise-wide use. Palantir CEO Alex Karp made similar arguments last week, criticizing the per-token approach that both Anthropic and OpenAI rely on while pointing to open-weight models as a more workable path for enterprise customers. "Something has gone completely wrong," Karp told CNBC 3

3

. Companies that once encouraged employees to use AI tools when costs were lower are now introducing usage caps, encouraging employees to use the right tool for each task, and adopting open-source models to manage expenses 4

4

.

Price War Emerges as Market Response

The good news for enterprise buyers is that a price war is already underway. DeepSeek has made a 75% discount permanent, and rivals are racing to match these lower prices 2

2

. A wave of startups is chasing cheaper inference costs to squeeze more output from every chip. Whether this adds up to Arora's demanded 90% reduction remains uncertain, since efficiency gains can be swallowed by ever-heavier usage. The opening for Chinese AI labs that charge less than U.S. companies due to more efficient models and lower energy costs adds another dimension to the competitive landscape 4

4

. For now, the message from a customer running a cybersecurity giant is clear: AI vendors' products remain too expensive to deploy everywhere enterprises want to use them. Coming from Palo Alto Networks, it's a signal model makers cannot ignore as they balance AI pricing needs to fall against their own AI infrastructure investments.